Modified added activation function based exponential robust random vector functional link network with expanded version for nonlinear system identification

This paper presents an improved non-iterative random vector functional link network hybrid model with better input-output representation, improved generalization for nonlinear dynamic system identification. The modified random vector functional link network model uses random weights between the input layer and hidden or enhancement layer neurons whose outputs are obtained by using two suitably weighted activation functions and additionally it provides a weighted direct link between the trigonometric functions based exponentially expanded the inputs and the output node. This novel architecture provides a direct link of inputs and its nonlinear expanded version in a higher dimensional space to the output node along with a randomized version of the hidden layer operating with an optimized added activation function to handle the chaotic nature of the non-linear dynamic systems. Also the weights and the parameters associated with the summed activation functions are optimized using an efficient modified sine cosine algorithm based monarch butterfly algorithm with levy distribution optimization algorithm with better exploitation and exploration capabilities in order to improve overall identification accuracy. To authenticate the efficiency of the proposed model, five benchmark dynamic plants are examined; the achieved outputs are compared with recognized methods like extreme learning machine, conventional random vector functional link network, and enhanced random vector functional link network with single activation function and least mean square. The method proposed here exhibits improved performance accuracy which is superior to the considered models. The proposed model is also compared with some iterative existing methods and found suitable by taking into consideration the merits of non-iterative approach,